The AI Revolution: How We Got Here
The first installment in a series on the sources of the present dilemma.
Like Swaine’s World? Scroll to the bottom of this post to tip the author or to subscribe.
Image of the Week
This was our river about ten years ago, August 2014. It look about the same now, allowing for seasonal variation of vegetation.
THE AI THREAT: HOW DID WE GET HERE?
The personal computer was born in a time of social ferment, when idealism ran high. Many of the people… were equally passionate about opening up the arcane technology of the computer to everyone. ‘Computer power to the people’ was their rallying cry….
That’s what we wrote in the Preface to the third edition of Fire in the Valley. Paul Freiberger and I were reporters in Silicon Valley covering a revolution from the front lines. We’d interview some kitchen-table entrepreneur about their fledgling company’s new product and dutifully take notes, and then we’d push our chairs back and say, “Now tell us about how you got started.” And they’d let their hair down and tell us the stories. And one thing that became obvious as we put all the individual stories together was that there were two interwoven rebellions going on in the 1960s and 70s.
One thread was political, involving resistance to the war in Viet Nam and the fight for civil rights and women’s rights. The rebellion invited a counter-rebellion, and in the 1960s and 70s America was a nation divided. The divisions led tragically to the Kent State massacre in 1970.
The other thread was less overt, but deep-seated: it involved a distrust, often an outright fear, of computer technology. We said:
Computers in the 1960s were massive…. Only government agencies, universities, and big businesses could afford to own a computer. And they were obscure and sinister, typically operated by a white-coated “priesthood” of specially-trained operators…. In the 1960s computers were widely viewed as a dehumanizing tool of the bureaucracy.
They evoked fear: fear that these giant brains would take our jobs, would out-think us, overthrow us, replace us. And the fear led to rebellion.
This second thread, the rebellion against the computer priesthood, was the driving force behind the personal computer revolution we wrote about in Fire in the Valley. Electronics hobbyists, model-train fanatics, model rocket builders, science fair winners — they were a motley group of nerds but they all rankled against the priesthood and they all began to see the promise of the semiconductor chips now becoming available. They could imagine having their own computer. Even building their own computer. And wresting the power of computers from the priesthood, from IBM and the other giant corporations.
Crazy.
But they did it.
That was the story we told in Fire in the Valley, and it was an upbeat, empowering story. Because the other thread of rebellion, the social rebellion against authority more generally, was also part of the revolution. “Computer power to the people” was the theme of Ted Nelson’s influential Computer Lib. Everybody read it — everybody involved in the revolution, that is. Steve Wozniak had Nelson’s book on his desk when he was working on the Apple II. Nelson was rightly called the Thomas Paine of the personal computer revolution.
The timeframe of the events detailed in the first edition of Fire in the Valley was basically 1974–1984. By 1984 the founders of the early personal computer and software companies were millionaires and their companies were listed on the NASDAQ. Things were changing. A revolution ended.
And today, with the nation divided as it hasn’t been in decades and with people genuinely worried that their job, even their entire profession, might shortly be replaced by computer technology, with experts predicting that we’re five years away from Artificial General Intelligence, calling the growth of AI capabilities an “existential risk” to humanity — it seems like we’ve gone full circle.
Or maybe it’s more of a spiral: Yeats’s “widening gyre” comes to mind. Not to mention “what rough beast.”
So how did we get here? I don’t mean the political divide, I won’t try to explain that. But the return to the Fear of the Machine. What exactly has brought that back?
Over the next few weeks, I plan to review the technological developments that led to the current fear of AI.
The most immediate and obvious triggering event, of course, was the release of Chat-GPT in November of 2022. There’s a lot more to the AI story than that, but let’s start there, with the company that made Chat-GPT, OpenAI.
The Chat-GPT Bombshell
The website proclaims:
“OpenAI is an AI research and deployment company. Our mission is to ensure that artificial general intelligence benefits all of humanity.”
“We are governed by a nonprofit and our unique capped-profit model drives our commitment to safety. This means that as AI becomes more powerful, we can redistribute profits from our work to maximize the social and economic benefits of AI technology.”
“We founded the OpenAI Nonprofit in late 2015 with the goal of building safe and beneficial artificial general intelligence for the benefit of humanity.”
Artificial General Intelligence, or AGI. Not AI. Not Generative AI, the current hot version of AI. AGI is the hypothesized ultimate AI, the AI that will match and exceed human intelligence, the AI that poses an existential risk to human survival. It would be the mission of OpenAI to head off that risk, to make sure that AGI would be safe and would benefit humanity rather than threaten it.
But when you develop technology, you have to productize it. So then there had to be the for-profit arm of the non-profit body:
“We designed OpenAI’s structure — a partnership between our original Nonprofit and a new capped profit arm — as a chassis for OpenAI’s mission….”
That was the concept behind OpenAI, the company behind Chat-GPT and several generations of GPT models. Quoting Techpedia: “OpenAI Inc. serves as the overarching non-profit umbrella, while the commercial aspirations are handled by OpenAI LP, its for-profit arm.”
How has that concept worked out? That could be the subject of a different article, and there are in fact plenty of such articles. Instead, here’s a fake video that might as a metaphor for how it’s working.
GPT Means Generative Pre-trained Transformer
Understanding the technology behind Chat-GPT means knowing about neural networks and large language models and reinforcement learning and the transformer model, specifically the generative pre-trained transformer, or GPT. I’ve talked about some of that in past posts and will probably get into them again in future posts. But for now, here’s the relevant history leading up to the Chat-GPT release:
In 2020, OpenAI had honed its GPT models to the point where they thought a public demo project might be warranted. From MIT Technology Review:
GPT-3 is a big leap forward. The model has 175 billion parameters (the values that a neural network tries to optimize during training), compared with GPT-2’s already vast 1.5 billion. And with language models, size really does matter.
GPT-3 could generate apparently meaningful text, responding appropriately to questions. It could write code that worked. It could also spit out falsehoods, drivel, and code that didn’t work. It wasn’t a finished product: it wasn’t a product at all, really, but a technology. But a chatbot based on GPT-3 would be a product.
They planned to release such a chatbot later that year.
But they didn’t.
Two years later, when GPT-3 had been superpassed by GPT-4, which was close to release, they hadn’t released that chatbot. But soon OpenAI executives began to worry that a competitor would release a chatbot first. Kevin Roose, in the New York Times: “So they decided to dust off and update an unreleased chatbot that used a souped-up version of GPT-3…. Thirteen days later, ChatGPT was born.”
OpenAI released to the public its ChatGPT, based on GPT-3.5, in November 2022. Within two months it had something like 30 million users, getting five million hits a day. It would be challenging to identify another software product that took off that quickly.
OpenAI’s developers were surprised by the response. I won’t say they were shocked, since their servers were apparently prepared for the five million hits per day. But they were clearly surprised by some of the criticism of ChatGPT.
Sandhini Agarwal, talking to MIT Technology Review:
We did find that it generated certain unwanted responses, but they were all things that GPT-3.5 also generates. So in terms of risk, as a research preview… it felt fine.
But users didn’t all feel fine about the unwanted responses. John Schulman:
I underestimated the extent to which people would probe and care about ther politics of ChatGPT. ChatGPT fails a lot…. I think we just have to be very up-front, and manage expectations, and make it clear this is not a finished product.
“Not a finished product.” True, and ChatGPT was also not a serious product. It was technology demo for people to play with and kick its tires. But other products using GPT technology are serious.
Agarwal:
The stakes right now are definitely a lot higher. One thing that obviously really matters with these models is the context they’re being used in. Like with Google and Microsoft, even one thing not being factual became such a big issue because they’re meant to be search engines. The required behavior of a large language model for something like search is very different than for something that’s just meant to be a playful chatbot.
But search engine companies like Google and Microsoft are plowing ahead with GPT technology in their search engines. And Gen Zers are turning to GPT products for mental health and career advice. The technology is permeating everything.
I come back to that phrase of John Schulman’s: “Not a finished product.” OpenAI’s developers, to their credit, are aware of and concerned about bad behavior in their software. Bias in training data. Jailbreaking. They understand that OpenAI has a mission to keep AI software safe, and I believe they take that responsibility seriously.
And the things they are doing do seem like the kinds of things that can make a product safer. But OpenAI and other AI companies are not just building products, they are bringing into existence new technologies, with capabilities that did not formerly exist. And it’s not obvious that they have any idea how to make those technologies safe.
I guess we’ll find out how it all works out.
BEFORE YOU GO…
The Pragmatic Bookshelf
These days my day job is editing books on programming for The Pragmatic Bookshelf. Nice folks, high quality books. Here’s what’s selling well on their site right now.
Blogroll
AI Supremacy
Ahead of AI
Mark Watson’s AI Books and Blog
Kent Beck’s advice for geeks
When We Were Trekkies
Tales from the Jar Side
Geena Davis Institute on Gender in Media
Bookshop.org
New York Review of Books
Pragmatic Bookshelf
ICYMI
Thanks for reading. You can read all the back issues of Swaine’s World at my blog home.
Coming Attractions
In the coming weeks I intend to dig deeper into the story of the AI revolution and how we got here.